A Data-Driven Framework for Optimizing Propranolol Dosage Using Support Vector Regression and Reinforcement Learning

The accurate prediction and adjustment of drug dosages requires precision to maximize therapeutic benefits while minimizing harm. This research attempts to model a hybrid machine learning framework combining Support Vector Regression (SVR) and Reinforcement Learning (RL) for individualized Propranolol dosage optimization using patient-specific clinical, enzymatic, and lifestyle data. A retrospective dataset comprising patient file, lifestyle indicators, and enzyme profile was used to train an SVR model for initial dosage prediction. Reinforcement Learning was subsequently applied to refine predictions through simulated feedback loops. Model performance was assessed using Mean Squared Error (MSE), R-squared (R²), and F1-score. Statistical comparisons between SVR predictions, RL-refined dosages, and physician-prescribed doses were performed using paired t-tests and one-way ANOVA. The SVR model achieved high predictive accuracy (MSE = 0.3554; R² = 0.9835), indicating its suitability for dosage estimation. The RL-refined model demonstrated a slight decrease in accuracy (MSE = 0.9928; R² = 0.9539). Statistical tests showed no significant improvement with RL (paired t-test: t = -1.1132, p = 0.2672; ANOVA: F = 0.0165, p = 0.9836). Mean predicted dosages across SVR, RL, and physician prescriptions were closely aligned (24.85 mg, 24.83 mg, and 24.93 mg, respectively). This study demonstrates that even standalone SVR may yield Propranolol dosage estimates with high accuracy, highlighting its prospective usefulness in clinical settings as a direct yet reliable tool for use in customized healthcare. While RL does offer some level of flexibility, the statistical value of improvements made was negligible, making RL beneficial but not necessarily critical. The proposed model shows that AI systems can aid in formulating evidence-based clinical judgments for dosing medications.